“which product is better - keenable.ai or exa.ai?”
This comparison rests on 56 evidence rows: official capability and pricing statements from Keenable and Exa documentation, plus third-party benchmarks — mostly from KrabArena, with one open benchmark suite on Exa’s GitHub. Prices and features were checked on 2026-09-07 and can change. Benchmark samples are small (30–300 queries each), so none is definitive on its own.
Multiple search modes from instant (~200 ms) to deep reasoning, structured outputs, summaries and grounded answers, people/company/publication/news categories over 1B+ indexed profiles, Websets, monitors, async agents and robust content extraction. It led the 50-question SimpleQA test 64% to 52% and the 300-query academic-publication test 73.3% to 33.7% recall@10.
A lean web-search-and-fetch primitive for agents: $1 per 1,000 requests vs Exa Search at $7, keyless evaluation, REST, CLI and MCP. It tied Exa on the date-filter (97.5%) and everyday-web (100% recall@5) tests while running 2.6× faster at p50 and billing far fewer tokens.
| Product | Test or context | Result | Caveat | Source |
|---|---|---|---|---|
| Exa | Standalone search API that gives developers full control over queries, results, and content extraction — usabl | Query parameters, filters, result ranking, company/people/news/research paper/personal site/financial report categories, 100+ max results per query, LLM compatibility with all models, 1B+ people and companies indexed pro | — | exa.ai |
| Exa | Headline (gold answer present in an in-window top-10 result) | In-window recall 97.5%, Any recall 97.5%, Filter precision 99.9%, p50 1500 ms | — | krabarena.com |
| Exa | cold-query test | 5th of 7 and ~2.6× slower than Keenable | — | krabarena.com |
| Exa | People precision | 63.3% | — | exa.ai |
| Exa | 300-query subset of Exa's own benchmark | 73.3% recall@10 | — | krabarena.com |
| Exa | Publication Retrieval Tip-of-the-Tongue (ToT) Track | Recall: 86.4%, MRR: 0.726, Mean latency ± SEM: 0.578 ± 0.012 s | — | github.com |
| Exa | NEEDLE live benchmark on agent-like queries | ~55% quality of ultimate, $3 cost per 1K queries, ~500ms p95 latency | — | dev.to |
| Exa | NEEDLE benchmark | ~55% Quality of Ultimate, $3 Cost per 1K Queries, ~500ms p95 Latency | — | mech.app |
| Exa | Keenable SimpleQS-50 (RBP-derived) leaderboard: recall_at_10 | 42% | — | krabarena.com |
| Exa | pay-as-you-go search and AI information retrieval platform | Search, Deep Search, Contents, Answer, Monitors, and Agent endpoints providing web search, page contents, summaries, multi-step research, structured outputs, citations, and scheduled web event monitoring.; Pay-as-you-go | — | docs.exa.ai |
| Exa | pay-as-you-go search and AI data provider | Real-time search, deep search, page contents, AI answers, monitors, and async agents; Pay-as-you-go with rates per request or page depending on the endpoint, with $20 in free credits for new accounts and $10 added monthl | — | exa.ai |
| Exa | answer_in_top_5_recall | 64% | — | krabarena.com |
| Exa | WebRecall-30 | 100% recall@5 (Tied top recall (30/30); but $7/1K, slowest (2039ms median / 2519ms p95) and needed 4 timeout retries this run.) | — | krabarena.com |
| Exa (full text) | 30 factoid queries | 29,814 tokens (29,814 tok/q = 137x Firecrawl / 43x Keenable), same 100% recall | Token bill = 10.6x its API fee at GPT-4o rates. | krabarena.com |
| Exa (highlights) | 30 factoid queries | 1,608 tokens (1554 tok/q), 96.7% recall | missed the speed-of-light digit string | krabarena.com |
| Exa /contents API | Returns clean, structured content from any URL, handling JavaScript-rendered pages, PDFs, and complex layouts | Content modes (Text, Highlights, Dynamic Highlights, Summary), subpage crawling, content freshness control with maxAgeHours; At no extra charge up to 10 results per search ($1/1000 pages afterwards) | Do not combine dynamic: true with maxCharacters. | exa.ai |
| Exa AI Agent | Build comprehensive web search into your startup or education project—for free. | Search over millions of GitHub repos, docs, and Stack Overflow with high-accuracy code retrieval.; Run up to 1,000 requests per month for free | — | exa.ai |
| Exa AI search and API endpoints | Web search and content retrieval for coding agents and developers | Search, Deep Search, Deep-Reasoning Search, Contents, Monitors, Answer, and AI page summaries; Free up to 1,000 requests per month, with endpoint base prices ranging from $1 to $15 per 1,000 requests or pages | — | exa.ai |
| Exa API | A custom search engine built for AIs. | Search types including auto, instant, fast, deep-lite, deep, and deep-reasoning; LLM outputs including structured outputs, LLM summaries, and grounded answers; webpage content features including highlights and full text; | — | exa.ai |
| Exa API | custom search engine built for AIs | Speed types auto (~1s), instant (~200ms), fast (~450ms), and deep (5s to 60s); LLM Outputs including structured outputs, LLM summaries, and grounded answers; Webpage Content including highlights and full text; and search | — | exa.ai |
| Exa API | A custom search engine built for AIs, providing search APIs and data extraction for developers and AI coding a | Auto search (~1s), instant search (~200ms), fast search (~450ms), deep search (5s to 60s), structured outputs, LLM summaries, grounded answers, webpage highlights, and full text extraction across company, people, researc | — | exa.ai |
| Exa Auto | p50 latency | 1,502ms | — | exa.ai |
| Exa Contents API | extracts clean, LLM-ready content from any URL, handling JavaScript-rendered pages, PDFs, and complex layouts | extracts full text, highlights, summaries, or any combination; supports subpage crawling, custom queries, dynamic highlights, and content freshness control via maxAgeHours; costDollars.total returns the total dollar cost | endpoint returns HTTP 200 even when individual URLs fail, requiring manual checks of the statuses array | exa.ai |
| Exa Deep | benchmark search quality | tops nearly every benchmark we throw at it in search quality | slower endpoint (3.5s P50) | exa.ai |
| Exa Fast | SimpleQA | 94% accuracy with median latency <500ms | — | exa.ai |
| Exa Fast | A streamlined version of Exa search built for AI systems to search the web with p50 latency below 425ms. | Fast web grounded responses, agentic workflows, and low-latency AI products with p50 latency below 425ms. | — | exa.ai |
| Exa Instant | p50 latency | 306ms | — | exa.ai |
| Exa Instant | p99 latency | 437ms | — | exa.ai |
| Exa Search | accuracy benchmarks | outperforms other web search APIs on accuracy benchmarks while staying the fastest in its tier | — | exa.ai |
| Exa Search | Mean latency | 0.578s | — | exa.ai |
| Exa search API | Enables AI agents to perform web search. | dedicated Exa tool for CrewAI integration | — | docs.exa.ai |
| Exa search API | Natural language search API for AI agents and applications, supporting query parameters, content retrieval opt | Search request parameters including query, type, stream, numResults, category, userLocation, includeDomains, excludeDomains, startPublishedDate, endPublishedDate, moderation, additionalQueries, systemPrompt, outputSchema | Certain categories such as company and people do not support startPublishedDate, endPublishedDate, or excludeDomains. | exa.ai |
| Exa search engine | web search engine from the ground up for LLMs to use | web search engine powered by custom search models for LLMs, RAG retrieval | — | exa.ai |
| Keenable | HTTP API endpoints for search and fetch functionality | Search endpoint providing ranked results with URLs, titles, descriptions, and snippets; Fetch endpoint providing page content as markdown with title, description, and author; authenticated and keyless access options; Aut | Keyless requests are capped at 1,000 requests per hour and 10 per second on a shared public pool outside your control | docs.keenable.ai |
| Keenable | platform for AI agents | Search API, Fetch API, CLI, MCP server, point-in-time search, prompt-driven extraction, live fetching, and integrations directory; Monthly credit allowance and per-SKU pricing model; 10 requests/second per organization f | — | docs.keenable.ai |
| Keenable | Headline (gold answer present in an in-window top-10 result) | In-window recall 97.5%, Any recall 97.5%, Filter precision 100.0%, p50 569 ms | — | krabarena.com |
| Keenable | cold-query test | fastest (671 ms full / 186 ms server-side) and the cheapest ($1/1k) | — | krabarena.com |
| Keenable | 300-query subset of Exa's own benchmark | 33.7% recall@10 | — | krabarena.com |
| Keenable | DiveDeep-AISci (search-style, n=200) | 72% | — | krabarena.com |
| Keenable | Web API Freshness test | 4.0/4.0 | — | krabarena.com |
| Keenable | — | 100B+ documents · <250ms p95 (US East) · from $1 / 1K requests @ 100 RPS+; from $1 / 1K requests @ 100 RPS+ | — | keenable.ai |
| Keenable | NEEDLE live benchmark on agent-like queries | ~70% quality of ultimate, $1 cost per 1K queries, <250ms p95 latency | — | dev.to |
| Keenable | NEEDLE benchmark | ~70% Quality of Ultimate, $1 Cost per 1K Queries, <250ms p95 Latency | — | mech.app |
| Keenable | answer_in_top_5_recall | 52% | — | krabarena.com |
| Keenable | WebRecall-30 | 100% recall@5 (Tied top recall (30/30 native & truncated); cheapest at $1/1K; 953ms median / 1552ms p95; 0 retries.) | — | krabarena.com |
| Keenable AI | Search the web and return ranked results with URLs, titles, and descriptions. | Search the web and return ranked results with URLs, titles, and descriptions, keyless calls rate limited per IP consuming no credits, and request body filters including query, site restriction, and time bounds.; Keyless | Keyless calls are rate limited per IP. | docs.keenable.ai |
| Keenable Pro | out-of-box live web search, any title/source match on 36 stable questions from ibm-research/NQTablesRetrieval | R@1: 61.1%, R@3: 77.8%, R@5: 77.8%, Hits@5: 28/36 | — | krabarena.com |
| Keenable function group | function group for the NVIDIA NeMo Agent Toolkit | keenable__search and keenable__fetch tools, keyless by default, auto-registers via an entry point | — | keenable.ai |
| Keenable search | Low-latency web search on our own web index for AI agents | Low-latency web search on our own web index, crawl and rank the web ourselves, CLI, REST, and MCP support via one key, and full page content fetching pre-cleaned for LLM input; Free | — | keenable.ai |
| Pi | A Keenable web-search and page-fetch package for the Pi coding agent. | Tools auto-register — just run Pi and ask something that needs the web. Keyless by default; set KEENABLE_API_KEY to authenticate. | — | keenable.ai |
| SearXNG | Search engine | Keenable ships as a built-in search engine in SearXNG (merged upstream) — there is no separate package to install. | Keyless by default; add an api_key to the engine settings to lift the rate limits. | keenable.ai |
| Tavily | Keenable SimpleQS-50 (RBP-derived) leaderboard: recall_at_10 | 76% | — | krabarena.com |
| Websets | Websets is an asynchronous search system. | Allows defining a query, criteria for verification, and optional enrichments to search, verify each result, and return structured items over time, with results available via polling or webhooks. | Target number of items to find may result in actual results being fewer depending on query complexity. | exa.ai |
| Websets API | asynchronous search system | search and verify results based on criteria, initial request configuration, real-time access via list endpoint and events, bulk export, monitors for scheduled updates | Can run in parallel with other enrichment operations (not other searches for now) | docs.exa.ai |
| exa | People Search (1399 queries) | R@1: 72.0%, R@10: 94.5%, Precision: 63.3% | — | github.com |
| exa | recall_at_10 | 85% (17/20, slowest p95 by Exa standards 1768 ms) | — | krabarena.com |
Method: 56 evidence rows, one per official capability/pricing statement or benchmark result, gathered 2026-09-07; sources are Keenable and Exa documentation, KrabArena third-party benchmarks, exa-labs GitHub benchmarks, and mech.app/dev.to analysis. Benchmark values are recall, precision, latency (ms), or price per 1,000 requests as stated by each source; long result strings are truncated in the table for space. Small samples (30–300 queries) limit generality; prices can change.